Detecting the Absence of Lung Sliding in Lung Ultrasounds Using Deep Learning

Round 1

Reviewer 1 Report

The authors present a deep-learning method to detect the absence of lung sliding motion in lung ultrasound. The field is new, and the authors present an interesting addition to this field. The methodological part is well-done. However, as it is very long the presentation of the method results a bit confusing. I would suggest adding a scheme to help the reader to follow the overall methodological flow.

I have only minor comments:

Introduction

• Lung ultrasound has been demonstrated to be successfully used for monitoring post-surgery complications not only from the authors of the present study. Please, also cite other groups who provided results in this direction, such as Smargiassi et al. Multidisciplinary Respiratory Medicine 2019, Galetin European Journal of Cardio-Thoracic Surgery 2020.
•  “Previous clinical trials performed by the co-authors of this paper showed that lung ultrasound (LUS) can be successfully used instead of X-ray” I suggest a more cautious phrasing, such as “Previous clinical trials showed that lung ultrasound could be used successfully as a primary imaging modality after non-cardiac thoracic surgery for diagnostics of pneumothorax and pleural effusion.”

Methods

• Could you specify the probe placing?
• Were the images from the lung not affected by the surgical procedure also analysed?

Segmentation Labels

1. How did you menage separate regions of the lung attached to the largest region?
2. Did you visually check the result of these segmentation?

Discussion

• Is there any effect of emphysema on the segmentation results?
• Did you see any difference in the segmentation results when the images are acquired by different operators?

Author Response

Dear reviewer, 

thank you very much for the review.

We added citation and rephrased the sentence from the introduction  to: “Previous clinical trials showed that lung ultrasound could be used successfully as a primary imaging modality after non-cardiac thoracic surgery for diagnostics of pneumothorax and pleural effusion.”

Methods

• Could you specify the probe placing?
• The physician scans the anterior and lateral sides of the thorax using a linear probe. The probe is placed between two ribs longitudinally to the body, and both lungs are examined. As for the precise placement of the probe on the rib cage our dataset does not contain this information.
• Were the images from the lung not affected by the surgical procedure also analysed?
• Yes.  The lung not affected by the surgical procedure serves as a comparison for the physician and these videos were included in the dataset.

Segmentation Labels

1. How did you manage separate regions of the lung attached to the largest region?
1. We leverage simple imaging processing techniques, we check the neighborhood of pixels and if they contain pixels with the same class, they are an interconnected region, then we measure the area of the region and compare it to the others.
2. Did you visually check the result of this segmentation?
1. Yes, we performed a visual check and in most cases, it performed well.

Discussion

• Is there any effect of emphysema on the segmentation results?
• We do not possess the dataset with emphysema LUS. However, some of the samples in our dataset are edge cases, where the whole LUS structure is fuzzy, therefore we expect it to work with emphysema too.
• Did you see any difference in the segmentation results when the images are acquired by different operators?
• Our dataset metadata does not contain operator identifiers. Therefore, we cannot say if there is a difference between operators in segmentation performance.

Reviewer 2 Report

Authors provided novel deep learning method for automated M-mode ultrasound classification. The obtained accuracy, sensitivity, and specificity of 89, 82, and 92% are good to be determined as M-mode ultrasound images.
The number of the samples for LSP and LSA are good for providing the data for deep learning techniques. Future work for 3D ultrasound images for CNN looks promising candidate for the proposed concept. However, some references are needed. Authors also need to add more comments for experimental data.Therefore, the manuscript can be minor revision and published with some correction if authors follow the guidelines.

1. Please use abbreviated journal names in the reference sections.
2. Please provide the city, country, and date information of the conference papers in the reference sections.
3. What represents the red mark in Figure 5 ?
4. In Line 4, X-raying -> X-ray.
5. Please provide the company information of "Sonoscape".
6. Please explain why authors selected ResNet 18 for classification.
7. In Table 1, precision, recall, and IoU for pleura looks lower than others. Is there any reason ?
8. In Figure 7, Mean difference for 256 f.ctr looks higher than others. If possible, please describe that.
9. Please provide the reference (The advantages of LUS include avoiding unnecessary radiation exposure,) with the reference (Kim, K. and Choi, H. (2021). High-efficiency high-voltage class F amplifier for high-frequency wireless ultrasound systems. PloS one, 16(3), e0249034.).
10. Please provide the reference (Pneumothorax is an abnormal collection of air in the pleural space.) with the reference (Heidecker, Jay, et al. "Pathophysiology of pneumothorax following ultrasound-guided thoracentesis." Chest 130.4 (2006): 1173-1184.) or another reference.

Author Response

Dear reviewer,

thank you very much for the review.

1. Please use abbreviated journal names in the reference sections.

• We abbreviated all the journals, where it was available.

2. Please provide the city, country, and date information of the conference papers in the reference sections.

• We added the city, country, and date information to every conference reference.

3. What represents the red mark in Figure 5 ?

• Red represents randomly erased areas. It represents the sample area that was randomly erased for the training of CNN.

4. In Line 4, X-raying -> X-ray.

• Done.

5. Please provide the company information of "Sonoscape".

• We added the company name into the section  Dataset description.

6. Please explain why the authors selected ResNet 18 for classification.

• The main reason is the repeatability of the experiments. While application of the custom network sometimes yields better results, we wanted to rely on well-tested architecture.

7. In Table 1, precision, recall, and IoU for pleura looks lower than others. Is there any reason?

• The lung tissue and rib shadows are well constrained. The lung tissue is constrained by the pleura and the anechoic shadows of the ribs, but the contours of the pleura tissue are fuzzier, therefore there is a difference in these metrics.

8. In Figure 7, Mean difference for 256 f.ctr looks higher than others. If possible, please describe that.

• Videos contain probe motion and angling, which is typical for clinical LUS videos. This may cause a decrease in the performance of our architecture because the time slices are quite long, in our case 256 frames last for 8.5 seconds, due to the motion, they contain a lot of noise.

Finally, we added both references.